Four kinds of statement
The Atlas makes four kinds of claim, and treats them differently.
- Editorial judgement. Ledes, plain-language explanations, guides and comparisons are written by the Atlas. They are opinions about emphasis and usefulness, not measurements, and they are written in the site's own voice.
- Measured facts. Population, area, elevation, coordinates, climate normals and every data series. These carry an observation period and a named producer, always.
- Grounded research. Sections marked deep dive. Written by a language model from one named source document that was fetched in full, with the source and the number of characters read printed on the page.
- Third-party reporting. News. Headline, publisher, date, short summary, link out. Never reproduced.
Population, area and elevation
These come from Wikidata claims. Where a claim carries a point-in-time qualifier, the Atlas takes the most recent qualified value and renders the year alongside it — so a figure reads "40,665 (2023)" and can be audited, rather than floating free. Wikidata is a compilation, not a primary source: for anything consequential, go to INSTAT, whose census and register data Wikidata is generally reflecting.
Albanian place figures need one further care. A figure for "Berat" may mean the city, the municipality (which includes surrounding villages) or the county (which includes four other municipalities). Every page states which unit it is describing.
Climate normals
Most "weather in X by month" tables on the web are uncited and undated. The Atlas computes its own. For every place with coordinates it fetches the complete daily record for 1 January 1994 to 31 December 2023 from the Open-Meteo historical archive, which serves ECMWF ERA5 reanalysis, and reduces it: mean daily maximum and minimum temperature, mean monthly precipitation total, the mean count of days with at least 1 mm of precipitation, and mean daily sunshine duration. Coastal places additionally get mean sea surface temperature from the Open-Meteo marine archive.
Missing days are dropped rather than zero-filled, because a zero would be a fabricated observation. The consequence is that Sarandë and Theth are measured on exactly the same basis and over exactly the same window, which is what makes them comparable. These are long-run averages, not forecasts; any individual year can differ substantially.
Economic and demographic series
From the World Bank Indicators API for the country code ALB. Each observation is stored with its period, its producer and the date the Atlas retrieved it. Values are never interpolated, carried forward or estimated: where the producer has published nothing, the row is blank and says so. Current-dollar series mix real growth with price and exchange-rate movement, which is why the growth series is given separately rather than derived here.
The deep-research pass
For each subject with an English Wikipedia article, the plain-text extract of the whole article is fetched. A language model is given that text and instructed to write structured sections using only what the text contains, in a defined house style, with a list of banned marketing phrases. Output that fails to parse is salvaged to its last complete section and then resampled; a page that cannot produce a usable result keeps whatever it had before.
Research is stored in a separate table from the curated text, so a re-run can never overwrite hand-written editorial content. The source article and the character count are shown in the Sources section of every page that has one.
The quality gate and noindex
Each entity is scored on the substance it actually carries: length of the plain-language
explanation, number of curated facts, number of published images, number of graph
relationships, whether it has twelve months of climate normals, and how many researched
sections it has. Pages below the threshold remain live and linked but are served
noindex and labelled as stubs. This is what prevents a programmatic reference site
from becoming a farm of thin pages.
Relationships
The "connected subjects" on every page are edges in a curated graph, not similarity matching. Each edge was written by hand as part of the entity record — near for places that are physically close, related for subjects that genuinely belong together, in for containment. Nothing on this site is a "you may also like".
What this method cannot do
It cannot verify a claim that appears in a single source. It cannot tell you the ferry timetable for next Tuesday. It cannot resolve genuinely contested questions — Illyrian continuity, religious affiliation percentages, the true number of bunkers — and where those arise it says they are contested rather than picking a side. And it cannot substitute for professional advice on anything legal, financial, immigration-related or medical.